Segmentation of anatomical structures in chest radiographs using supervised methods: a comparative study on a public database Revised version Bram van Ginneken, Mikkel B. Stegmann, and Marco Loog * 11th August 2004 * B. van Ginneken and M. Loog are with the Image Sciences Institute, University Med- ical Center Utrecht, Heidelberglaan 100, 3584 CX Utrecht, The Netherlands. E-mail: {bram,marco}@isi.uu.nl, URL: http://www.isi.uu.nl/. M.B. Stegmann is with Informat- ics and Mathematical Modelling, Technical University of Denmark, DTU, Richard Petersens Plads, Building 321, DK-2800 Kgs. Lyngby, Denmark. E-mail: [email protected], URL: http://www.imm.dtu.dk/. 1
47
Embed
Segmentation of anatomical structures in chest radiographs ... · The chest radiographs are taken from the JSRT database [5]. This is a publicly available database with 247 PA chest
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Segmentation of anatomical structures in chest
radiographs using supervised methods: a
comparative study on a public database
Revised version
Bram van Ginneken, Mikkel B. Stegmann, and Marco Loog ∗
11th August 2004
∗B. van Ginneken and M. Loog are with the Image Sciences Institute, University Med-
ical Center Utrecht, Heidelberglaan 100, 3584 CX Utrecht, The Netherlands. E-mail:
bram,[email protected], URL: http://www.isi.uu.nl/. M.B. Stegmann is with Informat-
ics and Mathematical Modelling, Technical University of Denmark, DTU, Richard Petersens
The task of segmenting the lung fields, the heart, and the clavicles
in standard posterior-anterior chest radiographs is considered. Three su-
pervised segmentation methods are compared: active shape models, ac-
tive appearance models, both first proposed by Cootes et al. [1] and a
multi-resolution pixel classification method that employs a multi-scale fil-
ter bank of Gaussian derivatives and a k-nearest-neighbors classifier. The
methods have been tested on a publicly available database of 247 chest
radiographs, in which all objects have been manually segmented by two
human observers.
A parameter optimization for active shape models is presented, and it
is shown that this optimization improves performance significantly. It is
demonstrated that the standard active appearance model scheme performs
poorly, but large improvements can be obtained by including areas outside
the objects into the model.
For lung field segmentation, all methods perform well, with pixel clas-
sification giving the best results: a paired t-test showed no significant
performance difference between pixel classification and an independent
human observer. For heart segmentation, all methods perform compara-
bly, but significantly worse than a human observer. Clavicle segmentation
is a hard problem for all methods; best results are obtained with active
shape models, but human performance is substantially better.
In addition, several hybrid systems are investigated. For heart segmentation,
where the separate systems perform comparably, significantly better performance
can be obtained by combining the results with majority voting.
As an application, the cardio-thoracic ratio is computed automatically
from the segmentation results. Bland and Altman plots indicate that all
methods perform well when compared to the gold standard, with confi-
dence intervals from pixel classification and active appearance modelling
very close to those of a human observer.
All results, including the manual segmentations, have been made pub-
licly available to facilitate future comparative studies.
Index terms — Chest radiographs, Segmentation, Lung field segmentation,
Heart segmentation, Clavicle segmentation, Active shape models, Active ap-
pearance models, Pixel classification.
Running title: Segmentation of anatomical structures in chest radiographs
2
1 Introduction
A large amount of literature in the medical image analysis research community
is devoted to the topic of segmentation. Many methods have been developed
and tested on a wide range of applications. Despite these efforts, or perhaps
because of the large number of algorithms that have been proposed, it remains
very difficult for a system designer to decide which approach is best suited for a
particular segmentation task. Fortunately, there is a growing awareness in the
medical image research community that evaluation and performance character-
ization of segmentation methods is a critical issue [2, 3]. Such evaluations are
greatly facilitated by the availability of public image databases with manual an-
notations on which researchers can test and compare different algorithms. For
this study, we have annotated a public database, and have made the manual
segmentations available [4].
We compare three methods for segmenting five important anatomical struc-
tures in the single most acquired medical image: the standard posterior-anterior
(PA) chest radiograph. To this end, these structures –the lung fields, the heart,
and the clavicles– have been segmented manually by two observers indepen-
dently in 247 radiographs from the publicly available JSRT (Japanese Society of
Thoracic Radiology) database [5]. The fact that each object has been manually
segmented twice allows one to use one manual segmentation as gold standard
and compare the performance of automatic methods with that of an indepen-
dent human observer. The web site of the annotated JSRT database [4] allows
other researchers to upload the results of other segmentation algorithms applied
to the database and we invite the medical image analysis research community
to do so.
Accurate segmentation of anatomical structures in chest radiographs is es-
sential for many analysis tasks considered in computer-aided diagnosis. These
include various size measurements, the determination of the presence of pul-
monary nodules or signs of interstitial lung disease. Knowledge about the lo-
cation of the clavicles can be used to reduce false positive findings or to detect
lesions hiding ‘behind a clavicle’ more reliably.
The methods considered here are active shape models (ASM) [6, 1], active
appearance models (AAM) [7] and pixel classification (PC). ASM is a popular
segmentation method, with many internal parameters. We consider how to tune
3
these parameters. AAM has recently found widespread application in medical
image segmentation. In this work we use an implementation available in the
public domain [8] and compare the standard AAM scheme with an extension
in which the surroundings of objects are modelled as well. PC is a classical
segmentation method, but the basic concept is so general that it can be imple-
mented in many different ways. We propose an implementation in which both
position and local image derivatives are used as input features and show how a
multi-resolution implementation and an approximate k-nearest neighbor classi-
fier lead to a relatively fast scheme that yields accurate segmentations. Finally,
we also consider three hybrid approaches. The first one fuses the results of the
best performing ASM, AAM and PC scheme by majority voting. The other
hybrid schemes uses a “tissue map” produced from the probability output of
the PC scheme as input for the ASM and AAM method, respectively.
Each of the methods examined here is supervised. This means that example
images with the desired output need to be supplied for training. This makes the
methods versatile; by supplying different training images and annotations, each
method can be applied to many different segmentation tasks, including the ones
investigated here. This is in contrast to rule-based schemes that are specifically
designed to handle one segmentation task.
The article is organized as follows. Section 2 briefly reviews previous work
on segmentation of lung fields, heart and clavicles in chest radiographs. Section
3 describes the data. The segmentation methods are presented in Section 4.
Section 6 presents the results, followed by a discussion in Section 6. Section 7
concludes.
2 Previous work
Segmentation of lung fields in PA chest radiographs has received considerable
attention in the literature. Rule-based schemes have been proposed by Li et
al. [9], Armato et al. [10], Xu et al. [11, 12], Duryea and Boone [13], Pietka
[14], and Brown et al. [15]. Lung segmentation by pixel classification using
neural networks has been investigated by McNitt-Gray et al. [16], and Tsujii
et al. [17]. Vittitoe et al. [18] developed a pixel classifier for the identification
of lung regions using Markov random field modeling. An iterative pixel-based
4
classification method related to Markov random fields was presented in [19].Van
Ginneken and Ter Haar Romeny proposed a hybrid method that combines a
rule-based scheme with a pixel classifier [20]. ASM has been used for lung field
segmentation in [21, 22].
Segmentation of the outline of the heart has been studied by several re-
searchers, usually with the aim of detecting cardiomegaly (enlarged heart size).
For this purpose, only parts of the heart border need to be known. Published
methods typically use rule-based schemes, using edge detection and a geomet-
rical model of the heart shape [23, 24, 25, 26, 27].
The segmentation of clavicles in chest radiographs has, to the best of our
knowledge, not been studied before.
3 Materials
3.1 Image data
The chest radiographs are taken from the JSRT database [5]. This is a publicly
available database with 247 PA chest radiographs collected from 13 institutions
in Japan and one in the United States. The images were scanned from films to
a size of 2048 × 2048 pixels, a spatial resolution of .175 mm/pixel and 12 bit
gray levels. 154 images contain exactly one pulmonary lung nodule each; the
other 93 images contain no lung nodules.
3.2 Object delineation
Each object has been delineated by clicking points along its boundary using a
mouse pointer device. These points are connected by straight line segments. For
the ASM and AAM segmentation methods, these contours need to be converted
to a fixed number of corresponding points. To this end several additional, dis-
tinguishable points on the contour are clicked by the user, indicating anatomical
or other characteristic landmarks. These characteristic points are assumed to
correspond. After the complete boundary has been defined, all but the corre-
sponding points are discarded and subsequent points are obtained by equidis-
tantly sampling a certain fixed number of points along the contour between the
aforementioned indicated points. This is illustrated in Fig. 1.
5
Two observers segmented 5 objects in each image. Observers were allowed
to zoom and adjust brightness and image contrast, and could take unlimited
time for segmentation. The first observer was a medical student, the second ob-
server a computer science student specializing in medical image analysis. While
neither of the observers were radiologists, both are familiar with medical images
and medical image analysis and have a good background in human anatomy.
Before the segmentations were made, both observers were instructed by an expe-
rienced radiologist until he was convinced that the segmentations produced by
the observers were reliable. After segmenting all objects, each observer reviewed
the results, and adjusted them to correct occasional errors and avoid bias due
to learning effects. When in doubt, they reviewed cases with the radiologist and
the radiologist provided the segmentation he believed to be correct. Review was
necessary in about 10% of all cases. Both observers segmented the images and
reviewed the results independently, but they did consult the same radiologist.
The segmentations of the first observer are taken as gold standard in this
study, to which the segmentations of a computer algorithm and the second
observer can be compared. The availability of a second observer allows for
comparisons between ‘human’ and ‘computer’ results.
3.3 Anatomical structures
In this work we consider the right and left lung, the outline of the heart and
the right and left clavicles1. It is important to carefully define what is meant
by the outline of an anatomical structure in a projection image.
The intensity in each pixel is determined by the attenuation of the radiation
by a column of body tissue. One could define the lung fields as the set of
pixels for which the radiation has passed through the lung fields. However, this
outline is impossible to determine from a frontal chest radiograph. Therefore we
adopt the following definition for the lung fields: any pixel for which radiation
passed through the lung, but not through the mediastinum, the heart, structures
below the diaphragm, and the aorta. The vena cava superior, when visible, is
not considered to be part of the mediastinum.
The heart is defined as those pixels for which radiation passes through the1Note that, by convention, a chest radiograph is displayed as if one is facing the patient.
This means that the right lung and clavicle are on the left in the image.
6
heart. From anatomical knowledge the heart border at the central top and
bottom part can be drawn. The great hilar vessels can be assumed to lie on top
of the heart.
For the clavicles, only those parts superimposed on the lungs and the rib
cage have been indicated. The reason for this is that the peripheral parts of the
clavicles are not always visible on a chest radiograph.
Fig. 1 shows one image and the annotated objects.
Figure 1: Left: the points indicated by the first observer on the first image of
the JSRT database to delineate lung fields, the heart, and the clavicles. The
anatomical or distinctive points are circled. The right lung contains 3 of these
points, the left lung 5, the heart 4 and each clavicle 6. Right: the landmarks
interpolated between the anatomical landmarks along the contours indicated on
the left for use in the ASM and AAM segmentation method. The total number
of landmarks is 166, with 44, 50, 26, 23 and 23 points in right lung, left lung,
heart, right clavicle, and left clavicle, respectively.
4 Methods
4.1 Active Shape Model segmentation
The following is a brief description of the ASM segmentation algorithm. The
purpose is mainly to point out the free parameters in the scheme; the specific
7
values for these parameters are listed in Table 1. Cootes et al. first introduced
the term active shape model in [28] and [6]. However, [6] does not include the
gray level appearance model and [28, 6] do not include the multi-resolution ASM
scheme. Both of these components are essential to obtain good segmentation
results with ASM in practice. Our implementation follows the description of
the ASM method given in [1] to which the reader is referred for details.
The ASM scheme consists of three elements: a global shape model, a local,
multi-resolution appearance model, and a multi-resolution search algorithm.
A set of objects in a training image is described by n corresponding points.
These points are stored in a shape vector x = (x1, y1, . . . , xn, yn)T . A set
of these vectors can be aligned by translating, rotating and scaling them so
as to minimize the sum of squared distances between the points (Procrustes
alignment, [29, 1]). Alignment can also be omitted, which will include the
variation in size and pose into the point distribution model which is subsequently
constructed. Let x denote the mean shape. The t principal components (modes
of variation in the shape model) of the covariance matrix of the shape vectors
are computed. The value of t is determined by fv, the amount of variation in
the training shapes one wants to explain. Shapes can now be written as
x = x + Φxbx, (1)
where Φx contains the modes of variation of the shape model and bx holds the
shape parameters. During the ASM search it is required to fit the shape model
to a set of landmarks. This is done by projecting the shape on the eigenvectors
in Φx and truncating each projection in bx to m times the standard deviation
in that direction.
A local appearance model is constructed for each landmark. On either side of
the contour at which the landmark is located, k pixels are sampled using a fixed
step size of 1 pixel, which gives profiles of length 2k + 1. Cootes et al. propose
to use the normalized first derivatives of these profiles [1]. The derivatives
are computed using finite differences; the normalization is such that the sum
of absolute values equals 1. Note that this requires a notion of connectivity
between the landmark points from which the direction perpendicular to the
contour can be computed.
As a measure for the goodness of fit of a pixel profile encountered during
8
search, the Mahalanobis distance to the set of profiles sampled from the training
set is computed. These profile models are constructed for Lmax resolutions. A
standard image pyramid [30] is used.
The search algorithm is a simple iterative scheme initialized by the mean
shape. Each landmark is moved along the direction perpendicular to the contour
to ns positions on either side, evaluating a total of 2ns + 1 positions. The
landmark is put at the position with the lowest Mahalanobis distance. After
moving all landmarks, the shape model is fitted to the displaced points, yielding
an updated segmentation. When a proportion pclose of points ends up within
ns/2 of its previous position, or when Nmax iterations have been made, the
search moves to the next resolution level, or ends. The highest resolution level
in our experiments was 256 × 256 pixels. The use of higher resolutions did not
improve performance.
In [1] suitable values are suggested for all parameters in the ASM scheme.
They are listed in Table 1 and they have been used in the experiments, referred
to as ‘ASM default’. In order to investigate the effect of different settings,
we performed pilot experiments on a small test set (to keep computation time
within bounds) and varied all settings within a sensible range, also given in
Table 1. The overall best setting was kept (last column in Table 1) and also
used in the experiments, referred to as ‘ASM tuned’.
4.2 Active Appearance Models
The active appearance model (AAM)segmentation and image interpretation
method [7] has recently received a considerable amount of attention in the image
analysis community [8]. AAM uses the same input as ASM, a set of training
images in which a set of corresponding points has been indicated.
The major difference to ASM is that an AAM considers all objects pixels,
compared to the border representation from ASM, in a combined model of
shape and appearance. The search algorithm is also different. This section
will summarize the traditional AAM framework, list the parameter settings and
describe alterations we applied for this segmentation task. Our implementation
was based on the freely available C++ AAM implementation described in [8].
An AAM is a generative model, which is capable of synthesising images of a
given object class. By estimating a compact and specific basis from a training
9
set, model parameters can be adjusted to fit unseen images and hence perform
both image interpretation and segmentation. The modelled object properties
are shape — using the shape vectors x — and pixel intensities (called texture),
denoted by t. As in ASM, variability is modelled by means of principal com-
ponent analyses (PCA). Prior to PCA modelling, shapes are Procrustes aligned
and textures are warped into a shape-free reference frame and sampled. Usu-
ally only the convex hull of the shape is included into the texture model. It is
also possible to model the inside of every closed contour. New instances for the
shape can be generated by Eq. 1, and, similarly, we have for the texture
t = t + Φtbt (2)
where t denotes the mean texture, Φt are eigenvectors of the texture disper-
sions (both estimated from the training set) and bt holds the texture model
parameters. To recover any correlation between shape and texture and obtain a
combined parameterization, c, the values of bx and bt are combined in a third
PCA,
WxΦTx (x− x)
ΦTt (t− t)
=
Wxbx
bt
=
Φc,x
Φc,t
c = Φcc. (3)
Here, Wx is a diagonal matrix weighting pixel distances against intensities.
Synthetic examples, parameterized by c, are generated by
x = x + ΦxW−1x Φc,xc
and
t = t + ΦtΦc,tc
and rendered into an image by warping the pixel intensities of t into the geom-
etry of the shape x.
Using an iterative updating scheme the model parameters in c can be fitted
rapidly to unseen images using the L2-norm as a cost function. See [1, 7] for
further details. As in ASM, a multi-resolution pyramid is used.
10
4.2.1 Parameter settings
This section lists the settings that were used in the AAM experiments. To
determine these settings, pilot experiments were performed. Our experience
suggests that the results are not sensitive to slight changes in these settings.
Segmentation experiments were carried out in a two-level image pyramid
(128 × 128 and 256 × 256 pixels). The use of coarser start resolutions was
investigated but did not improve performance.
The model was automatically initialized on the top level, by a sparse sam-
pling in the observed distribution of training set pose. This sparseness is ob-
tained by considering the convergence radius of each model parameter (inspired
by [1]), thus avoiding any unnecessary sampling. Since rotation variation was
completely covered by the convergence radius, no sampling was performed in
this parameter. From the training data, it was estimated that the model should
converge if initialized in a 2 by 2 grid around the mean position. Further, due to
the variation in size over the training set, each of these four searches was started
at 90%, 100%, and 110% of the mean size, respectively. Thus, 12 AAM searches
were executed in each image and the search producing the best model-to-image
fit was selected.
Both shape, texture and combined models were truncated at fv = .98, thus
including 98% of the variance. Bounds m on the combined eigenvalues were
three standard deviations. Model searches had a limit of 30 iterations at each
pyramid level. AAM parameter update matrices for pose and model parame-
ter were calculated using Jacobian matrices. These were estimated using every
15th training case. Parameter displacements were as follows: model parame-
ters: ±0.5σi, ±0.25σi (σi denotes the standard deviation of ith parameter), x-y
Table 2: Segmentation results for lungs, heart and clavicles, for each systemconsidered. All results are in terms of the overlap Ω, as defined in Eq. (4).The systems are ranked according to the median Ω. A paired t-test has beenapplied to each system and the system below it in this ranking. If the differenceis significant (p < 0.05), this is indicated with an asterix.
Table 3: Segmentation results for lungs, heart and clavicles, for each systemconsidered. All results are in terms of the mean absolute contour distance,given in millimeter. The systems are ranked according to the median meanabsolute contour distance. A paired t-test has been applied to each system andthe system below it in this ranking. If the difference is significant (p < 0.05),this is indicated with an asterix.
42
Lungs µ± σ min Q1 median Q3 maxASM GS 0.93 ± 0.03 0.72 0.92 0.94 0.94 0.96shape model fit 0.95 ± 0.02 0.76 0.94 0.95 0.96 0.97AAM GS 0.93 ± 0.02 0.84 0.93 0.94 0.94 0.96
Heart µ± σ min Q1 median Q3 maxASM GS 0.82 ± 0.08 0.50 0.77 0.82 0.88 0.95shape model fit 0.94 ± 0.04 0.44 0.94 0.95 0.96 0.98AAM GS 0.88 ± 0.05 0.66 0.86 0.89 0.92 0.96
Clavicles µ± σ min Q1 median Q3 maxASM GS 0.74 ± 0.13 0.19 0.71 0.78 0.81 0.90shape model fit 0.82 ± 0.06 0.35 0.80 0.83 0.85 0.90AAM GS 0.72 ± 0.08 0.26 0.68 0.74 0.78 0.88
Table 4: Segmentation results for lung, heart and clavicles, using the tunedASM system initialized with the gold standard, fitting the shape model fromthe ASM system directly to the gold standard, and the AAM whiskers systeminitialized with the gold standard. These results provide upper bounds for ASMand AAM systems. All results are in terms of the overlap Ω, as defined in Eq.(4).
Lungs µ± σ min Q1 median Q3 maxASM GS 2.18 ± 0.89 1.10 1.66 1.93 2.40 7.70shape model fit 1.56 ± 0.59 0.95 1.27 1.45 1.66 6.73AAM GS 1.93 ± 0.47 1.08 1.61 1.85 2.16 4.48
Heart µ± σ min Q1 median Q3 maxASM GS 5.87 ± 2.93 1.40 3.77 5.61 7.47 17.00shape model fit 1.67 ± 1.27 0.57 1.16 1.42 1.83 17.93AAM GS 3.61 ± 1.53 1.11 2.50 3.38 4.51 10.98
Clavicles µ± σ min Q1 median Q3 maxASM GS 1.99 ± 1.22 0.69 1.28 1.64 2.12 7.41shape model fit 1.28 ± 0.53 0.64 1.00 1.19 1.40 5.25AAM GS 2.05 ± 0.73 0.84 1.57 1.92 2.35 6.66
Table 5: Segmentation results for lung, heart and clavicles, using the tuned ASMsystem initialized with the gold standard, fitting the shape model from the ASMsystem directly to the gold standard, and the AAM whiskers system initializedwith the gold standard. These results provide upper bounds for ASM and AAMsystems. All results are in terms of the mean absolute contour distance, givenin millimeter.
43
Lungs
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Voting
PC post-
proce
ssed
Human PC
ASM/PC
AAM/PC
ASM tune
d
AAM whis
kers
+ BFGS
ASM defau
lt
AAM whis
kers
AAM defau
ltMea
n
Ove
rlap
Heart
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Human
Voting
ASM/PC
AAM/PC
AAM whis
kers
+ BFGS
PC post-
proce
ssed PC
ASM tune
d
ASM defau
lt
AAM whis
kers
AAM defau
ltMea
n
Ove
rlap
Clavicles
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Human
ASM tune
dVoti
ng
ASM defau
lt
ASM/PC
AAM whis
kers
+ BFGS
AAM/PC
AAM whis
kers
PC post-
proce
ssed PC
AAM defau
ltMea
n
Ove
rlap
Figure 4: Box plots of the overlap Ω for lungs, heart and clavicles for all methodsconsidered. The corresponding numerical values are listed in Table 2.
44
Lungs
0
2
4
6
8
10
PC post-
proce
ssed
Voting
Human
ASM/PC
AAM/PC
ASM tune
d
AAM whis
kers
+ BFGS PC
AAM whis
kers
ASM defau
lt
AAM defau
ltMea
n
Mea
n co
ntou
r dis
tanc
e
Heart
0
2
4
6
8
10
12
Human
Voting
PC post-
proce
ssed
ASM/PC
AAM whis
kers
+ BFGS
AAM/PC
ASM tune
d
AAM whis
kers PC
ASM defau
lt
AAM defau
ltMea
n
Mea
n co
ntou
r dis
tanc
e
Clavicles
0
2
4
6
8
10
Human
ASM tune
dVoti
ng
ASM defau
lt
ASM/PC
AAM whis
kers
+ BFGS
AAM/PC PC
PC post-
proce
ssed
AAM whis
kers
AAM defau
ltMea
n
Mea
n co
ntou
r dis
tanc
e
Figure 5: Box plots of the mean contour distance for lungs, heart and claviclesfor all methods considered. The corresponding numerical values are listed inTable 3.
Figure 6: Segmentation results for the gold standard, the second observer, thebest ASM, AAM, and PC systems, and the voting system which combines thethree latter systems, respectively. Four cases are shown ranging from easy(left) to difficult (right). See the text for details on how this ranking has beencomputed. Below each image the overlap Ω is listed for the right lung, left lung,heart, right clavicle and left clavicle, respectively.
46
0.35 0.4 0.45 0.5 0.55 0.6 0.65
−0.1
−0.05
0
0.05
0.1
0.15
0.2
0.25
Average CTR
Go
ld s
tan
da
rd −
Se
co
nd
ob
se
rve
r
0.35 0.4 0.45 0.5 0.55 0.6
−0.1
−0.05
0
0.05
0.1
0.15
0.2
0.25
Average CTR
Go
ld s
tan
da
rd −
AS
M t
un
ed
d = −0.00052; (-0.049,0.047) d = 0.012; (-0.058,0.083)
0.35 0.4 0.45 0.5 0.55 0.6 0.65
−0.1
−0.05
0
0.05
0.1
0.15
0.2
0.25
Average CTR
Go
ld s
tan
da
rd −
AA
M B
FG
S
0.35 0.4 0.45 0.5 0.55 0.6 0.65
−0.1
−0.05
0
0.05
0.1
0.15
0.2
0.25
Average CTR
Go
ld s
tan
da
rd −
PC
po
stp
roce
sse
d
d = 0.012; (-0.039,0.063) d = −0.0019; (-0.048,0.045)
Figure 7: Bland and Altman plots of the cardio-thoracic ratio computed from
the gold standard versus the second observer, ASM, AAM and PC. In the graphs
and below them the mean difference and the 95% confidence intervals (d−2σ, d+